agentharness-contrib

Provider implementations and common tools for AgentHarness. Each provider lives behind an optional extra, so the core stays dependency-free.

pip install "agentharness-contrib[anthropic]"

from agentharness import Agent, tool
from agentharness_contrib import AnthropicModel

@tool
def add(a: int, b: int) -> str:
    """Add two numbers."""
    return str(a + b)

agent = Agent(model=AnthropicModel("claude-opus-4-8"), tools=[add])
run = agent.run("What is 2 + 3?")
print(run.result)
run.trace.save("run.jsonl")   # real model call, fully recorded and replayable

AnthropicModel maps the core Model protocol onto Anthropic's Messages API. Pass client= to inject a preconfigured anthropic.Anthropic (or a test double); extra keyword arguments are forwarded to messages.create (e.g. thinking=..., system=...).

The Messages API translation is pure and SDK-free, so an agent's provider boundary is unit-tested without network calls.

Planned: OpenAI ([openai]) and Ollama / OpenAI-compatible providers.